Convex isoperimetric sets in the Heisenberg group
ROBERTOMONTI ANDMATTHIEURICKLY
Abstract. We characterize convex isoperimetric sets in the Heisenberg group.
We first prove Sobolev regularity for a certain class ofR2-valued vector fields of bounded variation in the plane related to the curvature equations. Then we show that the boundary of convex isoperimetric sets is foliated by geodesics of the Carnot-Carath´eodory distance.
Mathematics Subject Classification (2000): 49Q20 (primary); 53C17 (secondary).
1.
Introduction
We identify the Heisenberg group
H
1withC
×R
endowed with the group law (z, t)(z, t) =z+ z, t + t+ 2 Im (z¯z) ,
where t, t∈
R
and z= x + iy, z= x+ iy ∈C
. The Lie algebra of left-invariant vector fields is spanned byX = ∂
∂x + 2y∂
∂t, Y = ∂
∂y − 2x ∂
∂t and T = ∂
∂t,
and the distribution of planes spanned by X and Y , called horizontal distribution, generates the Lie algebra by brackets.
The natural volume in
H
1 is the Haar measure, which, up to a positive factor, coincides with Lebesgue measure inR
3 =C
×R
. Lebesgue measure is also the Riemannian volume of the left-invariant metric for which X , Y and T are orthonor- mal. We denote by|E| the volume of a (Lebesgue) measurable set E ⊂H
1. The horizontal perimeter (or simply perimeter) of E isP(E)=sup
E
Xϕ1+ Y ϕ2
d xd ydt ϕ1, ϕ2 ∈ Cc1(
R
3), ϕ21+ ϕ22≤ 1. (1.1) If P(E) < +∞, the set E is said to be of finite perimeter. Perimeter is left-invariant and 3-homogeneous with respect to the group of dilationsδλ:
H
1→H
1,δλ(z, t) = Received March 31, 2008; accepted October 14, 2008.(λz, λ2t), λ > 0, that is P(δλ(E)) = λ3P(E). Definition (1.1) of perimeter is mod- elled on De Giorgi’s notion of perimeter in Euclidean spaces which was generalized in [10] to Carnot-Carath´eodory spaces. For smooth sets (e.g. of class C2), perime- ter coincides with the Minkowski content and with the 3-dimensional Hausdorff measure of∂ E constructed by means of the standard Carnot-Carath´eodory metric in
H
1 (see [16] and [9], respectively). Volume and perimeter are related via the isoperimetric inequality|E| ≤ CisopP(E)4/3, (1.2)
where Cisop > 0 and E ⊂
H
1 is any measurable set with finite perimeter and volume. This inequality is proved by P. Pansu in [17] and [18] for smooth domains with the 3-dimensional Hausdorff measure of ∂ E inH
1 replacing P(E). In the general form (1.2), the inequality is proved in [10] (see also [8]). The problem of computing the sharp constant Cisop leads to the notion of isoperimetric set. A set E ⊂H
1 with 0 < |E| < +∞ is an isoperimetric set if it minimizes the isoperimetric ratioIsop(E) = P(E)4/3
|E| . (1.3)
Isoperimetric sets do exist: this is proved in [12] by a concentration-compactness argument. Pansu notes that the boundary of a smooth isoperimetric set has “constant mean curvature” and that a smooth surface has “constant mean curvature” if and only if it is foliated by horizontal lifts of plane circles with constant radius. Then he conjectures that an isoperimetric set is obtained by rotating around the center of the group a geodesic joining two points in the center. Recently, Pansu’s conjecture reappeared in [11].
The problem of determining isoperimetric sets is interesting for two reasons.
On the one hand, the non commutative group law makes it difficult to prove by a rearrangement argument that the isoperimetric ratio (1.3) is minimized by rotation- ally symmetric sets, as it is natural to conjecture. On the other hand, there is no regularity theory for measurable sets in
H
1minimizing perimeter with (or without) a volume constraint. So, new techniques and ideas are needed.All known results assume either symmetry or regularity (or both). In fact, assuming rotational symmetry and regularity it is easy to determine the isoperi- metric profile (see [13], and [22] for the general case). Actually, it suffices to assume the rotational symmetry of a certain horizontal section (see [7] and espe- cially the calibration argument in [21]). The solution of the rotationally symmetric case with no regularity assumption is in [14]. On the other hand, isoperimetric sets can be also determined assuming only the C2 regularity of the boundary and no symmetry (see [23]). Further evidence supporting Pansu’s conjecture is provided by [15], where a 2-dimensional version of the problem is solved. We refer to the monograph [3] for a more detailed introduction to the isoperimetric problem in the Heisenberg group.
In this article, we characterize isoperimetric sets which are convex. By convex set we mean a subset of
H
1 =R
3 which is convex with respect to the standardvector space structure of
R
3. Convexity is a left invariant property inH
1 because left translations are affine mappings.Theorem 1.1 (Convex isoperimetric sets). Up to a left translation and a dilation, any closed convex isoperimetric set in
H
1coincides withEisop=
(z, t) ∈
H
1 |t| ≤ arccos |z| + |z|1− |z|2, |z| ≤ 1
. (1.4)
The boundary of the set in (1.4) is foliated by Heisenberg geodesics, it is globally of class C2, but it fails to be of class C3at the north and south poles(0, ±π/2). At these points, the plane spanned by the vector fields X and Y , the horizontal plane, is tangent to the boundary.
We explain the main steps in the proof of Theorem 1.1. Let us consider a bounded convex set of the form
E =
(z, t) ∈
H
1| z ∈ D, f (z) ≤ t ≤ g(z), (1.5)
where D ⊂
R
2is a compact convex set in the plane with nonempty interior, and−g, f : D →
R
are convex functions. If E is isoperimetric, then the function f : D →R
satisfies the partial differential equationdiv
∇ f (z) + 2z⊥
|∇ f (z) + 2z⊥|
= 3P(E)
4|E| (1.6)
in int(D) \ ( f ). Here and in the following we let z = (x, y) and z⊥ = (−y, x).
We call the singular set( f ) =
z ∈ int(D) | − 2z⊥ ∈ ∂ f (z)
the characteristic set of f . This set is always contained in a line segment. The basic facts concerning
( f ) are discussed in the appendix. Equation (1.6) is derived in Section 2.
The curvature operator in (1.6) has been studied by several authors under dif- ferent regularity assumptions (besides the previous references, see also [4–6, 19, 20]). In the smooth case, the number
H = 3P(E)
4|E| (1.7)
is known as the (horizontal) curvature of∂ E. In our case, equation (1.6) is to be interpreted in distributional sense. The distributional derivatives of u(z) = ∇ f (z)+
2z⊥ are measures, because f is a convex function, and so the equation states that the distributional divergence of u/|u| is constant.
Our goal is to give equation (1.6) a pointwise meaning along integral curves of the vector field orthogonal to u/|u|. The first step is to show that the equation holds in the Sobolev sense. Precisely, we prove that the distributional derivative of u/|u|
is a measure which is absolutely continuous with respect to Lebesgue measure.
This result is a corollary of the following regularity theorem for BV vector fields, which is interesting in itself. We prove a slightly more general statement in the third section.
Theorem 1.2 (Improved regularity). Let ⊂
R
2be a bounded open set and let u∈ BV (;R
2) be a vector field. Suppose that:(i) There existsδ > 0 such that |u(z)| ≥ δ for
L
2-a.e. z∈ ;(ii) div u⊥ ∈ L1();
(iii) div u
|u|
∈ L1().
Then we have u/|u| ∈ W1,1(;
R
2).The improved regularity of the boundary is the starting point for the geometric characterization of convex isoperimetric sets. The vector fieldv(z) = −u⊥(z) = 2z−∇ f⊥(z) belongs to BVloc(int(D);
R
2). Moreover, its distributional divergence is in L∞, in factdivv = 4 in int(D). (1.8)
Thanks to the theory on the Cauchy Problem for BV vector fields recently devel- oped by Ambrosio in [1], the bound on the divergence ensures the existence of a unique regular Lagrangian flow : K × [− , ] → D starting from a compact set K ⊂ int(D), where > 0 is small enough. For
L
2-a.e. z ∈ K , the curve s → (z, s) is an integral curve of v passing through z at time s = 0. In Section 4, we show thatL
2-a.e. integral curve ofv is an arc of circle.Theorem 1.3 (Foliation by circles). Let E ⊂
H
1be a convex isoperimetric set of the form (1.5) and let K ⊂ int(D)\( f ) be a compact set. Then, forL
2-a.e. z ∈ K , the curve s → (z, s) is an arc of circle having radius 1/H, with H > 0 as in (1.7).The proof of Theorem 1.3 relies upon a reparameterization argument. The vector field v has a regular flow starting from K ⊂ int(D) \ ( f ) but it is only in BVloc(int(D);
R
2). On the other hand, v/|v| is in Wloc1,1(int(D);R
2), but its divergence is only in L1loc(int(D)) and so we have no regular flow for v/|v|. In order to compute the second order derivative of a generic integral curve of v, we introduce a suitable reparameterizationγ (s) = (z, τ(s)). We show that for any vector field w ∈ W1,1(;R
2) defined in some open neighborhood of K , the curveκ(s) = w(γ (s)) is in W1,1forL
2-a.e. z ∈ K and moreover˙κ = (∇w ◦ γ ) ˙γ in the weak sense. (1.9) Using the chain rule (1.9) with w = v/|v|, which is in W1,1 by Theorem 1.2, equation (1.6) can be given a pointwise meaning along the flow, and the integral curves ofv turn out to be arcs of circles.
Since the horizontal lift of the flow foliates the graph of the function f , it follows that the bottom (and upper) part of the boundary of E is foliated by geodesics of the Heisenberg Carnot-Carath´eodory metric. However, this is not yet
enough to finish our argument. We still need to show that the bottom and upper parts of∂ E match together in the proper way. To do this, we write E in the form
E=
(x, y, t) ∈
H
1| (y, t) ∈ F, h(y, t) ≤ x ≤ k(y, t), (1.10) for some compact convex set F ⊂
R
2 and convex functions h, −k : F →R
. The analysis carried out for f can be also carried out for h even though in this case computations are, unfortunately, more complicated. This provides the last piece of information needed to get the set Eisopin (1.4).A short overview is now in order. In Section 2, we derive the curvature equa- tions for convex isoperimetric sets. In Section 3, we prove the generalized version of Theorem 1.2 which is needed in Section 4, where we establish the foliation by geodesics property for convex isoperimetric sets. Finally, in Section 5 we prove Theorem 1.1. The results concerning convex sets are collected in the appendix at the end of the paper.
2.
Curvature equations for convex isoperimetric sets
We derive partial differential equations for certain vector fields built from the func- tions which parameterize the boundary of convex isoperimetric sets. We study graphs of the form t = f (x, y) and graphs of the form x = h(y, t). We use the following representation formula for perimeter. If E ⊂
H
1 is a bounded set such that∂ E is locally a Lipschitz surface, thenP(E) =
∂ E
(X · ν)2+ (Y · ν)2d
H
2, (2.1)whereν is a unit normal to ∂ E. Here and in the following, · denotes the standard inner product in
R
3 orR
2 (we think of X and Y as vectors inR
3).H
2is the 2- dimensional Hausdorff measure inR
3with respect to the usual Euclidean distance.For a proof of formula (2.1), see [9].
Graphs of the form t = f (x, y)
We denote elements of
H
1 =R
2×R
by(z, t) with t ∈R
and z = (x, y) ∈R
2. We write z⊥ = (−y, x). Let E be a convex set inH
1of the form (1.5). We define the characteristic set of the convex function f : D →R
as( f ) =
z ∈ int(D) | − 2z⊥ ∈ ∂ f (z)
, (2.2)
where∂ f (z) stands for the subdifferential of f at z.
Proposition 2.1 (Curvature equation I). Let E ⊂
H
1be a convex isoperimetric set with perimeter P(E) and volume |E|. Then the function f : D →R
satisfies in distributional sense the partial differential equationdiv
∇ f (z) + 2z⊥
|∇ f (z) + 2z⊥|
= H (2.3)
in int(D) \ ( f ), with H = 3P(E)/4|E|.
Proof. By Theorem A.1 in the appendix,( f ) is contained in the union of at most two line segments.
Letϕ ∈ Cc∞(int(D) \ ( f )) and for ε ∈
R
, consider the set Eε =(z, t) ∈
H
1| z ∈ D, f (z) + εϕ(z) ≤ t ≤ g(z) .We let P(ε) = P(Eε), V (ε) = |Eε| and R(ε) = P(ε)4/V (ε)3. If E is an isoperi- metric set, then R(ε) has a minimum at ε = 0, and then
R(0) = P3 V4
4PV − 3PV
ε=0
= 0. (2.4)
There existsε0 > 0 such that V(ε) = −
D
ϕ(z) dz for |ε| < ε0. (2.5) Letνεbe the exterior unit normal to∂ Eεand let Sε =
(z, f (z)+εϕ(z)) ∈
H
1| z ∈ Dbe the graph of f + εϕ. By the Heisenberg area formula (2.1) and from the standard area formula for graphs of functions in Euclidean spaces, we find
P(ε) = d dε
Sε
(X · νε)2+ (Y · νε)2d
H
2= d dε
D
|∇ f (z) + ε∇ϕ(z) + 2z⊥| dz.
(2.6)
By Proposition A.2 in the appendix, for any compact set K ⊂ int(D) \ ( f ), there isδ > 0 such that |∇ f (z) + 2z⊥| ≥ δ for
L
2-a.e. z ∈ K . Then we can interchange derivative and integral for all smallε. At ε = 0 we obtainP(0) =
D
∇ f (z) + 2z⊥
|∇ f (z) + 2z⊥| · ∇ϕ(z) dz. (2.7) From (2.4), (2.5) and (2.7) we find
4|E|
D
∇ f (z) + 2z⊥
|∇ f (z) + 2z⊥|· ∇ϕ(z) dz = −3P(E)
Dϕ(z) dz for arbitraryϕ ∈ Cc∞(int(D) \ ( f )). This is (2.3).
Graphs of the form x = h(y, t)
We denote points in
H
1 =R
×R
2 by(x, ζ) with x ∈R
andζ = (y, t) ∈R
2. Let E be a convex set inH
1of the form (1.10). We define the characteristic set of the convex function h : F →R
as the set(h) of the points ζ ∈ F such that the horizontal plane spanned by the vector fields X and Y at the point p= (h(ζ ), ζ) ∈R
×R
2is a supporting plane for E (i.e. the plane does not intersect the interior of E). By Theorem A.1 in the appendix,(h) is contained in the union of at most two line segments.Proposition 2.2 (Curvature equation II). Let E ⊂
H
1be a convex isoperimetric set with perimeter P(E) and volume |E|. Then the function h : F →R
satisfies in distributional sense the partial differential equation(∂y− 2h∂t) u1
|u|
− 2y∂t
u2
|u|
= H (2.8)
in int(F)\(h), with H = 3P(E)/4|E| and u = (u1, u2) = (hy−2hht, 1−2yht).
Proof. Notice that u ∈ BVloc(int(F);
R
2) and moreover, by Proposition A.3 in the appendix, for each compact set K ⊂ (int(F) \ (h)), there exists δ > 0 such that|u| ≥ δ
L
2-a.e. in K . Hence u/|u| ∈ BVloc(int(F) \ (h);R
2).Letϕ ∈ Cc∞(int(F) \ (h)) and for any ε ∈
R
consider the set Eε =(x, ζ) ∈
H
1| ζ ∈ F, h(ζ) + εϕ(ζ) ≤ x ≤ k(ζ ) .We let P(ε) = P(Eε), V (ε) = |Eε| and R(ε) = P(ε)4/V (ε)3. Denoting by Sε =
(h(ζ) + εϕ(ζ ), ζ) ∈
H
1 | ζ ∈ Fthe graph of h+ εϕ and by νεthe exterior unit normal to∂ Eε, from the Heisenberg area formula (2.1) and from the standard area formula we get
P(ε)= d dε
Sε
(X · νε)2+ (Y · νε)2d
H
2= d dε
F
1− 2y(ht+εϕt)2
+
(hy+εϕy) − 2(h+εϕ)(ht+εϕt)21/2 dζ.
Because we have(hy− 2hht)2+ (1 − 2yht)2≥ δ2> 0
L
2-a.e. in a neighbourhood of spt(ϕ), we can differentiate under the integral sign and we obtainP(0) =
F
(hy− 2hht)(ϕy− 2(ϕh)t) − 2yϕt(1 − 2yht)
(hy− 2hht)2+ (1 − 2yht)2 dζ
atε = 0. An integration by parts yields P(0) = −
F
ϕ d
(∂y− 2h∂t) u1
|u|
− 2y∂t
u2
|u|
,
where the derivatives of u/|u| are measures. If E minimizes the isoperimetric ratio then, as in the proof of Proposition 2.1, we get
4|E|
F
ϕ d
(∂y− 2h∂t) u1
|u|
− 2y∂t
u2
|u|
= 3P(E)
F
ϕ dζ
for arbitraryϕ ∈ Cc∞(int(F) \ (h)). This is (2.8).
3.
Improved regularity of the boundary
In this section, we prove a regularity result for vector fields with bounded variation in the plane arising from the parameterization of the boundary of convex isoperi- metric sets.
Let ⊂
R
2be an open set and let a, b ∈ C() be continuous functions. We consider the differential operatorM
acting on BV(;R
2) defined byM
u= (∂1− a∂2)u1+ b∂2u2. (3.1) In general,M
u is a measure in. We haveM
= div when a = 0 and b = 1.When a = 2h and b = −2y,
M
is the operator appearing in the left hand side of (2.8).We recall that a vector field u ∈ L1(;
R
2) has approximate limit ¯u(q) ∈R
2 at the point q ∈ iflimr↓0−
B(q,r)|u − ¯u(q)| d
L
2= 0. (3.2)Here, B(q, r) denotes the open Euclidean ball of radius r centered at q. The ap- proximate discontinuity set of u is the set Su of points in at which u has no approximate limit. The jump set of u is the set Ju of points q ∈ for which there exist u+(q), u−(q) ∈
R
2andνu(q) ∈ S1such that u+(q) = u−(q) andlim
r↓0−
B±(q,r)|u − u±(q)| d
L
2= 0, (3.3)where B±(q, r) =
q ∈ B(q, r) | ± (q− q) · νu(q) > 0
. Finally, the precise representative u∗ : →
R
2of u is:u∗(q) =
¯u(q) q ∈ \ Su,
1 2
u+(q) + u−(q)
q ∈ Ju,
0 q ∈ Su\ Ju.
(3.4)
By the Lebesgue density theorem, it is u= u∗
L
2-a.e. in.Theorem 3.1 (Improved regularity). Let ⊂
R
2be a bounded open set, let u= (u1, u2) ∈ BV (;R
2), and let a, b ∈ C() be continuous functions such that b = 0 in . Assume that:(i) There existsδ > 0 such that |u| ≥ δ
L
2-a.e. in;(ii)
M
u⊥∈ L1(), where u⊥= (−u2, u1);(iii)
M
u|u|
∈ L1().
Then u/|u| ∈ W1,1(;
R
2) and there exists a function µ : Ju → (0, +∞), such that u− = µu+on Ju.Proof. If u ∈ BV (;
R
2) the measure Du has the decomposition Du = Dau+ Dsu with DauL
2and Dsu ⊥L
2. Dau= ∇uL
2is the absolutely continuous part of the measure, with∇u ∈ L1(; M2×2) and M2×2denotes the space of 2× 2 matrices with real entries. Moreover, it is Dsu = Dcu + Dju, where Dcu = Dsu \ Su is the Cantor part and Dju = Dsu Su is the jump part. By the representation theorem of Federer–Vol’pert and by Alberti’s rank one theorem, the measures Dcu and Dju admit the representationsDju= (u+− u−) ⊗ νu
H
1 Ju and Dcu= η ⊗ ξ |Dcu|, (3.5) where |Dcu| is the total variation of Dcu and η, ξ : → S1are suitable Borel maps. The measure|Dcu| is absolutely continuous with respect toH
1. The set Ju is anH
1-rectifiable Borel subset of SuwithH
1(Su\ Ju) = 0. For these and related results on BV vector fields we refer the reader to the monograph [2].We claim that both the jump and the Cantor parts of D(u/|u|) vanish. Let F:
R
2→R
2be a smooth mapping such thatF(0) = 0, F(q) = q
|q| for |q| ≥ δ
2 and |∇ F| ∈ L∞(
R
2). (3.6) If|q| > δ/2, the derivative of F is∇ F(q) = 1
|q|3q⊥⊗ q⊥. (3.7)
By (i), the vector fieldv = F ◦ u is in BV (;
R
2) (this fact is implicitly assumed in (iii)).By the chain rule for BV vector fields, we have
Dav = ∇ F(u∗)∇u
L
2 and Dcv = (∇ F(u∗)η) ⊗ ξ |Dcu|. (3.8) A point q ∈ belongs to Jv if and only if q ∈ Ju and F(u+(q)) = F(u−(q)).Moreover, it is Djv =
F(u+) − F(u−)
⊗ νu
H
1 Ju. Notice that|u±(q)| ≥ δ for all q ∈ Ju, by (i). Then, the jump set ofv isJv =
q∈ Ju | u+(q)/|u+(q)| = u−(q)/|u−(q)|
(3.9) and the jump part of Dv is
Djv = u+
|u+| − u−
|u−|
⊗ νu
H
1 Ju. (3.10)By assumption (ii), the part of the measure−∂1u2+ b∂2u1+ a∂2u2concentrated on Ju vanishes. From the formula for Dju in (3.5), we can compute Dkjul for k, l = 1, 2, and we obtain
(u−2 − u+2), b(u+1 − u−1) + a(u+2 − u−2)
· νu= 0 (3.11)
H
1-a.e. on Ju.By assumption (iii), the part of the measure(∂1− a∂2)(u1/|u|) + b∂2(u2/|u|) concentrated on Jv vanishes. Thus, using (3.10), we can compute Dkj(ul/|u|) for k, l = 1, 2, and we get
u+1
|u+|− u−1
|u−|,au−1 − bu−2
|u−| −au+1 − bu+2
|u+|
· νu = 0 (3.12)
H
1-a.e. on Jv.From (3.12) and (3.11), we deduce that there existsλ ∈
R
such that the fol- lowing system of equations is satisfiedH
1-a.e. on Jv:
u+1
|u+| − u−1
|u−| = −λ(u+2 − u−2) bu+2 − au+1
|u+| − bu−2 − au−1
|u−| = λ
b(u+1 − u−1) + a(u+2 − u−2) . By elementary linear algebra, using b = 0, we obtain the equivalent system
u+1
|u+| − u−1
|u−| = −λ(u+2 − u−2) u+2
|u+| − u−2
|u−| = λ(u+1 − u−1)
⇔
u+
|u+|− u−
|u−|
·
u+− u−
= 0,
that is u−· u+ = |u−||u+|,
H
1-a.e. on Jv. It follows that u− = µu+H
1-a.e. on Jvfor someµ : Jv→ (0, +∞), and then u+/|u+| = u−/|u−|H
1-a.e. on Jv. This proves thatH
1(Jv) = 0 and thus Djv = 0.By assumption (ii), the Cantor part of the measure−∂1u2+ b∂2u1+ a∂2u2 vanishes. From Dcu= η ⊗ ξ |Dcu|, we can compute Dckul for k, l = 1, 2, and we find
(−η2, bη1+ aη2) · ξ = 0 (3.13)
|Dcu|-a.e. on . By assumption (iii), the Cantor part of the measure (∂1− a∂2)(u1/|u|) + b∂2(u2/|u|)
vanishes, too. Thus, lettingϑ = (ϑ1, ϑ2) = ((u∗)⊥ ⊗ (u∗)⊥)η, we can use (3.7) and (3.8) to compute Dck(ul/|u|) for k, l = 1, 2, and we find
ϑ1, bϑ2− aϑ1
· ξ = 0 (3.14)
|Dcu|-a.e. on . From (3.13) and (3.14), we deduce that there exists λ ∈
R
suchthat
ϑ1= −λη2
bϑ2− aϑ1= λ
bη1+ aη2
⇔
ϑ1 = −λη2
ϑ2 = λη1. Here we used b = 0. This, in turn, is equivalent with
ϑ · η =
(u∗)⊥⊗ (u∗)⊥ η
· η = 0
|Dcu|-a.e. on . Using the identity
(u∗)⊥ ⊗ (u∗)⊥ η
· η =
(u∗)⊥· η2
, we deduce that(u∗)⊥· η = 0, and thus
(u∗)⊥⊗ (u∗)⊥
η = 0 as well. By (3.7), this proves that(∇ F(u∗)η) ⊗ ξ = 0 |Dcu|-a.e. on , whence Dcv = 0 by (3.8).
Theorem 3.1 has the following corollaries.
Corollary 3.2. Let E ⊂
H
1 be a convex isoperimetric set and let f : D →R
be the function in (1.5). Then we have∇ f (z) + 2z⊥
|∇ f (z) + 2z⊥| ∈ Wloc1,1(int(D) \ ( f );
R
2). (3.15) Proof. The vector field u(z) = (u1(z), u2(z)) = ∇ f (z) + 2z⊥ satisfies div u⊥ =−4 in int(D). By Proposition A.2 in the appendix, |u| ≥ δ > 0 on a compact set in int(D) \ ( f ). Moreover, by Proposition 2.1, we have div(u/|u|) = 3P(E)/4|E|
in int(D) \ ( f ). The claim follows from Theorem 3.1 with a = 0 and b = 1.
Corollary 3.3. Let E ⊂
H
1be a convex isoperimetric set and let h : F →R
be the function in (1.10). Then we haveu
|u| ∈ Wloc1,1(int(F) \ ({y = 0} ∪ (h));
R
2), (3.16) with u= (u1, u2) = (hy− 2hht, 1 − 2yht).Proof. We use Theorem 3.1 with a = 2h and b = −2y. The vector field u is in B Vloc(int(F);
R
2) and satisfiesM
u⊥= (∂y−2h∂t)(2yht−1)−2y∂t(hy−2hht) = 2ht(1+2yht) ∈ L∞loc(int(F)).From Proposition A.3 in the appendix, it follows that u/|u| ∈ BVloc(int(F)\(h)), and moreover, by Proposition 2.2, u/|u| satisfies
M
u|u|
= (∂y− 2h∂t) u1
|u|
− 2y∂t
u2
|u|
= 3P(E) 4|E|
in int(F) \ (h). Now the claim follows from Theorem 3.1.
4.
Integration of the curvature equation
In this section we solve equations (2.3) and (2.8). These equations can be integrated along a regular Lagrangian flow and the solutions are suitable arcs of circle.
We recall the definition of regular flow in the plane. Let ⊂
R
2be an open set and let u ∈ BV (;R
2) ∩ L∞(;R
2). For q ∈ and > 0 small enough defineCq([− , ]; )
=
γ ∈ C([− , ]; ) γ (s) = q +
s
0
u(γ (σ)) dσ, s ∈ [− , ]
.
Let K ⊂ be a compact set. An
L
2-measurable map : K → C([− , ]; ) is a Lagrangian flow starting from K relative to u if(q) ∈ Cq([− , ]; ) forL
2- a.e. q ∈ K . With abuse of notation, is identified with the map : K ×[− , ] →, (q, s) = (q)(s). The flow is said to be regular if there exists a constant m ≥ 1 such that
1
m
L
2(A) ≤L
2((A, s)) ≤ mL
2(A) (4.1) for allL
2-measurable sets A⊂ K and for all s ∈ [− , ].Theorem 4.1 (Foliation by circles I). Let E ⊂
H
1 be a convex isoperimetric set and let f : D →R
be the function in (1.5). Let K ⊂ int(D) \ ( f ) be a compact set and ⊂ int(D) \ ( f ) an open neighborhood of K .For a sufficiently small > 0, there exists a (unique) regular Lagrangian flow
: K × [− , ] → of the vector field v(z) = 2z − ∇ f⊥(z). Moreover, for
L
2-a.e. z ∈ K , the integral curve s → (z, s) is an arc of circle with radius 4|E|/3P(E) oriented clockwise.For a convex isoperimetric set of the form (1.10) there is an analogous state- ment.
Theorem 4.2 (Foliation by circles II). Let E ⊂
H
1 be a convex isoperimetric set and let h : F →R
be the function in (1.10). Let K ⊂ int(F) \{y = 0} ∪ (h) be a compact set and ⊂ int(F) \
{y = 0} ∪ (h)
be an open neighborhood of K . For a sufficiently small > 0, there exists a (unique) regular Lagrangian flow : K × [− , ] → of the vector field v(ζ ) =
1− 2yht, 2yhy − 2h , ζ = (y, t) ∈ F. Moreover, for
L
2-a.e.ζ ∈ K , the projection onto the xy-plane of the curves → (h((ζ, s)), (ζ, s)) ∈
H
1, s ∈ [− , ], (4.2) is an arc of circle with radius 4|E|/3P(E) oriented clockwise.The existence of a (unique) regular Lagrangian flow stated in Theorems 4.1 and 4.2 follows from Ambrosio’s theory on the Cauchy problem for BV vector fields.
Theorem 4.3 (Ambrosio). Let ⊂
R
2be an open set and assume that: (i) u∈ BV (;R
2) ∩ L∞(;R
2);(ii) div u∈ L∞().
Then, for any compact set K ⊂ and for a small enough > 0, there exists a (unique) regular Lagrangian flow : K × [− , ] → starting from K relative to u.
The existence statement is Theorem 6.2 and the uniqueness statement is in [1, The- orem 6.4]. We do not need the uniqueness of the flow in our argument. Ambrosio’s theory holds more generally for non autonomous vector fields in any space dimen- sion.
The characterization of the flow in Theorems 4.1 and 4.2 is obtained on proving that a suitable reparameterization γ of a generic integral curve of the flow has a second order derivative which satisfies the equation ¨γ = −H ˙γ⊥ in a weak sense with H = 3P(E)/4|E|.
In order to make this reparameterization argument precise, let us consider an open set ⊂
R
2 and a vector field v in satisfying the assumptions of Theo- rem 4.3, that isv ∈ BV (;
R
2) ∩ L∞(;R
2) and div v ∈ L∞(). (4.3) The functionv is defined pointwise, i.e. we choose a representative in the equiva- lence class ofv. Our results hold independently of this choice. However, there is an exceptional set of points which may a priori depend on the representative.Given a compact set K ⊂ and a sufficiently small > 0, there exists a unique regular Lagrangian flow : K × [− , ] → starting from K relative to v. Let λ : →
R
be a measurable function such that0< c1≤ λ ≤ c2
L
2-a.e. in. (4.4) Then, forL
2-a.e. q∈ K , the curve s → λ((q, s)) is measurable andc1≤ λ((q, s)) ≤ c2 for a.e. s ∈ [− , ].
This follows from (4.1) by a Fubini-type argument. In fact, if N ⊂ is an
L
2- negligible set, then −1(N) ⊂ K × [− , ] is also negligible. Thus, forL
2- a.e. q ∈ K , the change of parameter σq : [− , ] → [σq(− ), σq( )] defined byσq(s) =
s
0 λ((q, ξ)) dξ
is bi-Lipschitz, strictly increasing and admits therefore a bi-Lipschitz, strictly in- creasing inverseτq : [σq(− ), σq( )] → [− , ], which satisfies
˙τq(s) = 1
˙σq(τq(s)) = 1
λ((q, τq(s))) for a.e. s ∈ [σq(− ), σq( )]. (4.5)
Consequently, for
L
2-a.e. q∈ K , the curve γq : [σq(− ), σq( )] → defined byγq(s) = (q, τq(s)) (4.6)
is absolutely continuous and satisfies
˙γq(s) = v(γq(s))
λ(γq(s)) for a.e. s∈ [σq(− ), σq( )], (4.7) i.e.γq is an integral curve of the vector fieldv/λ.
Theorem 4.4 (Chain rule for integral curves). Let ⊂
R
2be an open set, K ⊂be a compact subset, w ∈ W1,1(;
R
2) and > 0 small enough. ForL
2-a.e. q∈ K , the curvew ◦ γq withγq defined in (4.6) belongs to W1,1((σq(− ), σq( ));R
2) and its weak derivative is(∇w ◦ γq) ˙γq.Proof. Let{wk}k∈Nbe a sequence of smooth vector fieldswk ∈ W1,1(;
R
2) con- verging tow in W1,1(;R
2). The map (q, s) → w((q, s)) is measurable and integrable on K × [− , ]. By Fubini’s theorem and by the bounded volume dis- tortion property (4.1) of the flow, we have
K
− |wk((q, s)) − w((q, s))| ds dq ≤m
−
(K,s)|wk(q) − w(q)| dq ds, and
K
− ∇wk((q, s)) − ∇w((q, s))
v((q, s))ds dq
≤ v∞
−
K
|∇wk((q, s)) − ∇w((q, s))| dq ds
≤ mv∞
−
(K,s)|∇wk(q) − ∇w(q)| dq ds.
The curve s→ wk((q, s)) is Lipschitz, for all k ∈
N
and forL
2-a.e. q∈ K , with derivative s → ∇wk((q, s))v((q, s)). Passing to a subsequence and relabelling if necessary, we conclude thatklim→∞wk((q, ·)) = w((q, ·)) in W1,1((− , );
R
2)for
L
2-a.e. q∈K , and that the weak derivative of w((q,·)) is ∇w((q,·))v((q,·)).Combining this observation, (4.5) and (4.7), for anyϕ ∈ Cc∞((σq(− ), σq( ));
R
2)we get,
σq( )
σq(− )w(γq(s)) ˙ϕ(s) ds =
− w((q, s)) ˙ϕ(σq(s)) ˙σq(s) ds
= −
− ∇w((q, s))v((q, s)) ϕ(σq(s)) ds
= −
σq( )
σq(− )∇w(γq(s)) ˙γq(s) ϕ(s) ds.
Hence w ◦ γq ∈ W1,1((σq(− ), σq( ));
R
2), and its weak derivative is (∇w ◦ γq) ˙γq.Proof of Theorem 4.1. Let f : D →
R
be the convex function in (1.5). We con- sider the vector fields v(z) = 2z − ∇ f⊥(z) and u(z) = v⊥(z) = ∇ f (z) + 2z⊥ which are both in BVloc(int(D);R
2) ∩ L∞loc(int(D);R
2). We havedivv = 4 + fyx− fx y= 4. (4.8) If K ⊂ int(D) \ ( f ) is a compact set and
int(D) \ ( f ) is a suitable open neighborhood of K , the existence of a regular Lagrangian flow : K × [− , ] →for some > 0 is guaranteed by Theorem 4.3.
The functionλ : →
R
,λ = |v|, satisfies 0 < c1≤ λ ≤ c2L
2-a.e. in by Proposition A.2 in the appendix. By Corollary 3.2, the vector fieldw = v/λ belongs to W1,1(;R
2), and div w⊥= HL
2-a.e. in by (2.3), with H = 3P(E)/4|E|.Let γz : (σz(− ), σz( )) → be the integral curve of the vector field w defined in (4.6) and satisfying (4.7). We claim that γz parameterizes an arc of circle with curvature H . Since(z, ·) is a reparameterization of γz, proving the claim concludes the proof of Theorem 4.1. The following identities hold a.e. in (σz(− ), σz( )) for
L
2-a.e. z∈ K :(i) |w ◦ γz| = 1;
(ii) (w · ∂xw) ◦ γz= (w · ∂yw) ◦ γz = 0;
(iii) (div w⊥) ◦ γz = H.
Here, we are using the fact that if F, G : →
R
areL
2-measurable functions such that F = GL
2-a.e. in, then it is F ◦ = G ◦L
3-a.e. in K × [− , ], which is a consequence of (4.1).By Theorem 4.4, it isγz∈ W2,1((σz(− ), σz( ));
R
2) with ¨γz= (∇w ◦γz) ˙γz. Using (ii) and (iii), we compute¨γz· ˙γz⊥= (−div w⊥) ◦ γz= −H.
Moreover, by (i) we also have ¨γz· ˙γz = 0 a.e. in (σz(− ), σz( )). Then ¨γz= −H ˙γz⊥
a.e. and this implies thatγzis of class C∞. The claim follows.
Proof of Theorem 4.2. Let h : F →
R
be the convex function in (1.10). We use the variablesζ = (y, t) ∈ F. The vector field v : F →R
2v(ζ ) =
1− 2yht, 2yhy− 2h
, (4.9)
is in BVloc(int(F);
R
2) ∩ L∞loc(int(F);R
2), anddivv = −4ht− 2yht y+ 2yhyt = −4ht ∈ L∞loc(int(F)). (4.10) The vector fieldv is the projection onto the yt-plane of the vector field
(y, t) → (hy− 2hht)∂x+ (1 − 2yht)∂y+ (2yhy− 2h)∂t, (4.11) which is both horizontal and tangent to the graph of h at
H
2-a.e. point. We denote by u = (hy − 2hht, 1 − 2yht) the projection of the vector field (4.11) onto the x y-plane. The relation betweenv and u isv =
0 1
2y −2h
u. (4.12)
In (4.12), we think ofv and u as column vectors. The function λ : →
R
,λ = |u|, satisfies 0< c1≤ λ ≤ c2L
2-a.e. in by Proposition A.3 in the appendix.Let K and be as in the statement of Theorem 4.2. The existence of a regular Lagrangian flow : K × [− , ] → of v follows from Theorem 4.3.
Let F ∈ C∞(
R
2;R
2) be a mapping which satisfies (3.6) (with c1in place of δ). We consider the vector fields v/λ and w = F ◦ u in . Then w = u/|u| a.e. in. We have w ∈ W1,1(;
R
2) by Corollary 3.3. Observing that 2yhy − 2h = 2y(hy− 2hht) + 2h(2yht− 1), we also get that v/λ ∈ W1,1(;R
2). We claim that∇wv
λ = −Hw⊥
L
2-a.e. in, (4.13)with H = 3P(E)/4|E|. Indeed, using (3.7), we compute
∇wv
λ = (∇ F ◦ u) ∇uv λ= 1
|u|4
u⊥⊗ u⊥
∇u v
= 1
|u|3
u⊥· (∇u v) w⊥.
On the other hand, by (2.8) we have
M
w = H in , whereM
is the differential operator appearing in (3.1) with a= 2h and b = −2y. Let B be the 2 × 2 matrixB=
1 0
−2h −2y
. Using (3.7) we find
M
w = tr(∇ F ◦ u)∇uB
= 1
|u|3tr
(u⊥⊗ u⊥)∇uB
= 1
|u|3u⊥· (u⊥∇uB),
where, by a short computation based on (4.12), we have u⊥· (u⊥∇uB) = −u⊥· (∇u v). This ends the proof of (4.13).
Denote byγζ : [σζ(− ), σζ( )] → the integral curve of v/λ defined in (4.6). Then the curve s → w(γζ(s)) belongs to W1,1((σζ(− ), σζ( ));
R
2) forL
2-a.e.ζ ∈ K , and its weak derivative is equal to (∇w ◦ γζ) ˙γζ, by Theorem 4.4.Identity (4.13) holds along
L
2-a.e. curveγζ. Hence, by Theorem 4.4 applied tow, we haved
ds(w ◦ γζ) = −H(w⊥◦ γζ) (4.14) in the sense of weak derivatives.
The projection of the vector fieldζ = (y, t) → (h(ζ), ζ ) onto the xy-plane belongs to W1,1(;
R
2). Denote by κζ : [σζ(− ), σζ( )] →R
2 the projec- tion of the curve (h(γζ), γζ) onto the xy-plane. This projection is a reparame- terization of the curve in (4.2). We claim that κζ parameterizes an arc of cir- cle with curvature H oriented clockwise. By Theorem 4.4, forL
2-a.e. ζ ∈ K we have κζ ∈ W1,1((σζ(− ), σζ( ));R
2), and a short computation shows that the weak derivative of κζ is ˙κζ = w ◦ γζ. From (4.14), we deduce that κζ ∈ W2,1((σζ(− ), σζ( ));R
2) and ¨κζ = −H ˙κζ⊥. This implies thatκζ is of class C∞ and the claim follows.5.
Characterization of convex isoperimetric sets
In this section, we use Theorems 4.1 and 4.2 to characterize convex isoperimetric sets. We also use Theorem A.1 in the appendix on the structure of the character- istic set of bounded convex sets. Here, we recall some known facts about Carnot- Carath´eodory geodesics in the Heisenberg group.
An absolutely continuous path γ : [0, 1] →
H
1 is said to be horizontal if˙γ(s) ∈ spanR{X(γ (s)), Y (γ (s))} for a.e. s ∈ [0, 1]. The curve γ = (γ1, γ2, γ3) is horizontal if and only if
γ3(s) = γ3(0) + 2
s
0
˙γ1γ2− ˙γ2γ1
dσ. (5.1)
The plane curveκ = (γ1, γ2) is the horizontal projection of γ . If κ = (γ1, γ2) is a given plane absolutely continuous curve, then a curveγ = (κ, γ3) with γ3given by (5.1) for someγ3(0) is called a horizontal lift of κ. The standard sub-Riemannian length of the horizontal curveγ is
Length(γ ) =
1
0
|˙κ(s)| ds, (5.2)
whereκ is the horizontal projection of γ and |˙κ| is the Euclidean length of ˙κ ∈